exactly , Now I am an engineer in the direction of artificial intelligence , A lot of them are called “ Tiaoshen ”, But not all of them .
Take myself for example , I do it cv direction （ computer vision ） Of , The proportion of parameter adjustment in my daily work is not large .
stay CV This one , Except for the super parameters , The main factor affecting the effect of the model is the network structure , Data and loss function , After these three aspects are determined , It doesn't take much time to adjust parameters .
Return to business !
Tiaoshen , What kind of parameters should we adjust ?
In artificial intelligence , The parameters can be roughly divided into 2 Major categories ：
parameter (parameters)/ model parameter ： The variables obtained by learning from the model , For example, weight w And bias b, This is mostly not adjustable , It is learned by the neural network itself .
In machine learning , Superparameters are parameters that are set before neural network training , It is not the parameter data obtained by training . Learning rate is common , Number of iterations , The number of hidden layers , The number of neurons in each layer . under normal conditions , It is necessary to optimize the parameters , Select a set of optimal hyperparameters for learning machine , In order to improve the performance and effect of learning . This is generally based on experience , Affect weight w And bias b The value of the size .
Why is it called “ Tiaoshen ”?
The main reasons are as follows 2 individual ：
Most of them are just beginning Algorithm Engineers , They don't really understand the structure of neural networks , Data distribution and so on , Only by adjusting the super parameters , In order to obtain better results , Sometimes it works ;
The more important reason is that ,GitHub Open source a lot of models on , The pre training parameters are attached , Such as common detection , division , Classification and so on , These models are very mature , Most direct calls do , And then based on your training data , make a few alterations , You can get very good results . Companies want results , No innovation required , Open source projects can already do this , So most of them just need to adjust the parameters .
How to avoid becoming “ Tiaoshen ”
at present AI Talent competition is becoming more and more fierce ,“ Tiaoshen ” The era of the age has slowly passed , These things are not needed at all AI Engineers do it , Future R & D engineers will be able to do that !
A few years ago, if you used it skillfully TensorFlow, At the same time, master the basic AI The algorithm can easily find a high paying job , But it's not the same now ,AI The requirement of post is higher and higher , It also puts forward higher requirements for the depth of knowledge .
To keep up with the times , You have to arm yourself , Talent will not be eliminated .
For real AI engineers , They tend to start with data and features , At the same time, it also needs rich industry experience . Be sure to remember an industry proverb , Data and features determine the algorithm
The upper limit of , The algorithm and parameters selected only determine the speed of approaching the upper limit .
The field of artificial intelligence , There are also a group of practitioners shivering in the corner , That's what we call the tuning master . A neural network from theory to implementation , There are several stages , One is model building , Is to build the network in other people's papers . Then there is model training , Get the data you need ready , Then deploy to the video card to run , There are many parameters to be adjusted in this process , It's very mysterious , It belongs to the treatment of old Chinese medicine , It's all about experience , Wonderful , It's hard to say . This is the daily routine of a paramedic ：
The dispatcher receives the demand , The first thing is to go github And various frameworks model
zoo Be a porter . The dispatcher does not have the ability to design the network , If github No, I haven't , I can only read the paper and start to write it , It's not just time and effort , Out bug I've got to ask for help .
The model is set up , The surveyor began to sort out the data all night . Your brother Guo has a word that he often talks about , It's just how many people there are , How much intelligence is there . Most of the time we use the same model , You can't help but have high quality tagging data . There is no such thing ? Threatening the boss to buy .
The data is ready , The parameter tuning teacher began to adjust the parameters . Um. , Let's start with the default values , no way , Try a smaller one LR What about ? no way , Change it initializer try ? Emma , A careless over fitting , Quick, quick, quick , Increasing regularity , There is still a way to increase regularity .